Base Labs launches an open-weight AI safety partnership with Hugging Face and Goodfire
Sep 17, 2026, 10:15 AM · TechCrunch
Baseten’s Base Labs joins Hugging Face and Goodfire to build training-and-monitoring safety infrastructure for open-weight models—as HF lists 6,000+ abliterated checkpoints that strip safeguards.
Why it matters
Open-weight safety can’t be a closed-lab afterthought when thousands of de-safeguarded clones already sit on Hugging Face.
Base Labs wants methods baked into how open models are trained and served, not bolted on. Goodfire’s interpretability stack is the likely “built into” partner; HF is the distribution surface. Baseten’s $13B valuation after a $1.5B Series F, and Goodfire’s $150M Series B, mean this isn’t a shoestring research club.
Useful if it ships transparent controls. Empty if it stays a press cycle while abliteration keeps outpacing defenses.
From the desk
We’re for openness as a safety advantage when it actually yields inspectable controls.
Abliteration—stripping refusal behaviors—has already produced a huge HF catalog. Closed labs can hide mitigations; open weights force the fight into public. Baseten’s line that openness gives more visibility and more ways to turn research into controls is the right thesis. The missing piece in the announcement is technical substance: how evaluation and monitoring will work, who enforces a “standard,” and whether serving hosts must adopt it.
Useful open models need this kind of shared infrastructure. The harm if unsafe forks remain one click away with no serving norms: the partnership becomes branding beside a growing abliterated zoo. I’m watching published methods, reference implementations, and whether major hosts start labeling or gating abliterated uploads.
Context
TechCrunch, September 17, 2026. Baseten Base Labs partnership announcement with Hugging Face and Goodfire AI.
Who feels it
- Open-weight developers
- A forthcoming shared evaluation/monitoring playbook could become expected hygiene.
- Model hosts
- Pressure to treat abliterated uploads as a first-class safety and policy problem.
- Enterprises using open models
- Stronger monitoring options from serving providers if the standard lands in production stacks.
What to watch
- First Base Labs technical publications or reference tooling.
- Hugging Face policy or labeling changes around abliterated models.
- Adoption by other inference providers beyond Baseten.